Abstract: The forest fire smoke identification algorithm based on convolutional neural networks has some limitations, such as rather complicated smog feature extraction structure, tedious multi-scale feature fusion method, high computational complexity, limited application scenarios, high hardware requirements and difficult to adapt to the changing forest environment, which limit its application on the forest-fire prevention. To solve these problems, we proposed a lightweight convolutional neural network-based algorithm for detecting forest fire smoke. Firstly, based on re-parameterization technology and Cross Stage Partial Network, a lightweight structure for extracting smoke features was proposed to enhance the efficiency and velocity of smoke feature extraction. Secondly, a lightweight multi-scale smoke feature fusion method was developed utilizing the simplified Feature Pyramid Network and Path Aggregation Network to efficiently integrate smoke features across different scales. Then, a post-processing method for smoke detection was proposed, which involved the addition of similar smoke images to train the algorithm model, to mitigate the impact of non-fire smoke images and variations in application scenarios on detection accuracy. Finally, the algorithm was validated using the constructed smoke image dataset. The experimental results demonstrate that compared with other algorithms, the proposed algorithm is more accurate and faster with F₁ up to 82.6%, AP value up to 54.5%, and the maximum detection speed up to 869 images per second.